fix(otel): end all recording spans on non-terminal invocations

[ayushiahjolia]

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

Issue Link, if available

#624

Description

• Deferred the Workflow span in both plugins so it's started and ended in a single call on the terminal invocation, instead of being started every invocation and dropped un-ended on PENDING/RETRYING.
• Deferred operation spans in ExecutionOtelPlugin to onOperationEnd (retaining only a deterministic context between start and end), so a suspended operation no longer leaves a recording span abandoned.
• Force-end any still-open attempt spans at invocation end in ExecutionOtelPlugin so no recording span is ever abandoned.
• Carried the trace's sampling flags onto the synthetic Workflow/operation parent contexts in both plugins, so parent-based samplers drop children of an unsampled trace instead of forcing them to record as orphans.
• Retained operation start timestamps in ExecutionOtelPlugin as a fallback for virtual FLAT map/parallel ops, whose onOperationEnd reports null timestamps, so their spans start at the real start time.
• Added executionStartTime to InvocationEndInfo and populated it in onInvocationEnd, so plugins read the Workflow span start time from the end hook instead of caching it across hooks.

Demo/Screenshots

N/A

Checklist

• I have filled out every section of the PR template
• I have thoroughly tested this change

Testing

Unit Tests

Have unit tests been written for these changes? Yes, updated

Integration Tests

Have integration tests been written for these changes? N/A

Examples

Has a new example been added for the change? (if applicable) N/A

[github-actions]

Codex AI review

[P1] Do not assume an opaque agent provider is sampled

In the recommended no-arg Java-agent path, the provider is commonly hidden behind a classloader wrapper, making sdkTracerProvider null regardless of its sampler. Marking the synthetic Workflow parent sampled causes ExecutionOtelPlugin operations to export under a parent-based sampler even when the terminal Workflow root is dropped, producing orphan spans. Bridge the actual sampling result through the auto-configuration/classloader boundary and add a wrapped-provider test with a root-dropping sampler.

[github-actions]

Codex AI review

[P2] Reuse the Workflow root's exact sampling result

This probe uses empty attributes, while the real Workflow span is sampled with durable.execution.arn; the terminal span also invokes the sampler a second time. Attribute-based or stateful/rate-limiting samplers can therefore give the synthetic context and real root different decisions, and sampler-updated trace state is discarded. Query with the real attributes and preserve the complete SamplingResult through a sampler wrapper/cache so the terminal root reuses it. Add attribute-dependent and stateful sampler tests.

[github-actions]

Codex AI review

Two sampling-consistency regressions need correction before merge. Current tests cover only visible, deterministic samplers, missing the Java-agent and stateful-sampler failure modes.

Reviewed commit b865c79b2287d93efb7d317355685a578a33a458. Workflow run